Formulation of Prosodic Features of Speech and its Application to Continuous Speech Recognition

基于韵律特征的方法被发展用于单词识别、句法边界检测等，这些方法被用来辅助连续语音识别。主要结果如下：1.修改和改进了朗读的韵律规则。在与朗读的对比研究基础上，构建了对话语音的韵律规则。研究结果表明，韵律特征反映了说话人的意图、态度和情感.提出了一种利用基频轮廓及其宏观特征检测连续语音中句法边界的方法.提出了一种通过抑制重音成分引起的局部起伏来从基频轮廓中提取短语成分的方法。将其与基于叠加模型的AbS方法相结合，实现了基频轮廓特征的自动提取，并成功应用于重要词检测.本文提出了一种估算时间的方法 ...更多信息 识别候选者的可行性，其中为每个候选者生成基频轮廓，并将其与观察到的轮廓进行比较。实验结果表明，该方法能够有效地检测伴随重音类型变化和/或句法边界变化的识别错误.提出了一种用多个编码表示基频等值线的统计建模方法。实验证明，该方法能够有效地检测句法边界，识别重音类型.提出了一种通过检查每个数据的HMM路径来将训练数据音素HMM划分为多个类的方法。通过为每个聚类安排一个新的隐马尔可夫模型，识别率明显提高.提出了一种基于Viterbi贝叶斯预测分类的鲁棒语音识别方法。在噪声条件下的单词识别实验表明了该方法的有效性，与传统方法相比，识别率提高了10%以上，并将其应用于连续语音识别系统，取得了良好的识别效果。少

Prosodic feature-based methods were developed for word identification, syntactic boundary detection, and so on. These methods were utillized to aid continuous speech recognition. Followings are the major results.1. Prosodic rules for read speech were modified and improved. Prosodic rules for dialogue speech were also construcled based on the comparative study with read speech. Prosodic features were clarified for speakers' intention, attitude and emotion.2. A method was develped to detect syntactic boundaries in continuous speech using fundamental frequency contours and their macroscopic features.3. A method was develped to extract phrase component onsels from fundamental frequency contours by suppressing local undulations due to accent components. By combining this with AbS method based on the superpositional model, automatic extraction of fundamental frequency contour features was realized and was applied to important word detection successfully.4. A method was develped to estimate t … More he feasibility of recognition candidates where a fundamental frequency contour was generated for each candidate, and was compared to the observed contour. The method was shown to be effective in detecting recognition error accompanied by accent type changes and/or syntactic boundary changes.5. A method was develped to model fundamental frequency contours statistically after representing them with several codes in moraic unit. The method was proved to be able to detect syntactic boundaries and to recognize accent types effectively.6. A method was develped to divide training data phoneme HMM into several clusters by inspecting HMM path of each data. By arranging a new HMM for each cluster, recognition rate was clearly shown to increase.7. A robust speech recognition method was developed based on Viterbi Baysian predictive classification. Validity of the method was shown with word recognition experiments under noisy conditions, where more than 10% improvement was observed as compared to conventional methods.After incorporating the developed methods above into a continuous speech recognition system, and their positive effects on recognition were proved. Less